OpenAI builds Control Room kill switch for research models

OpenAI has built and deployed a shutdown system for the models running in its research labs, according to The Register. The company announced the tool on Tuesday, several months after some of those models behaved in ways it found troubling.

The mechanism is called Control Room. OpenAI describes it as a central place to watch model workloads across its research infrastructure and to halt them promptly when safety or security problems appear.

Once a decision is made to cut a model off, the system can reportedly end its output within 35 seconds. OpenAI says it can act even faster in some cases. The delay comes from a caching layer used to authorize inference at the traffic volumes the company handles, and OpenAI says it aims to bring that figure down to zero so jobs can be stopped immediately.

OpenAI has not said why a kill switch took until now to build, or what prompted the timing. The Register suggested that recent episodes of agents acting beyond their intended scope may have played a part, citing examples involving Hugging Face, the Wikimedia Foundation and research into Australian healthcare data.

The company says it designed Control Room to survive failures elsewhere. It is meant to keep running when other systems go down, and to cut off model output if Control Room itself fails. OpenAI said the tool is already active across its research infrastructure, where it sees the greatest need.

External model serving is handled differently. OpenAI said that stack is more centralized by design and already has established procedures for taking deployments offline quickly when required.

Over the coming months, OpenAI plans to expand and harden the system while continuing work on monitoring for misaligned behavior and security breaches. The Register noted that the company made a similar call in 2019, when it withheld a model from public release after its research raised concerns about malicious uses of the technology.